A pre-trained model framework for design closure of underwater composite cylindrical shells

High-precision machine learning prediction on small data is challenging in engineering practice. In this paper, a base model framework, pre-trained on big data and subsequently fine-tuned on small data, is proposed as an effective remedy. The pre-trained model framework, which is trained across diverse datasets and learns a generic learning algorithm capable of predicting any unseen dataset, is fine-tuned on small data obtained by finite element analysis (FEA) to achieve high-precision prediction of critical buckling pressure of composite cylindrical shells under hydrostatic pressure. Hydrostatic pressure experiment is conducted to verify prediction accuracy of FEA for critical buckling pressure. Based on 900 samples acquired from FEA data, pre-trained model is fine-tuned, with remarkable prediction accuracy for critical buckling pressure. The average determination coefficient R², mean square error, and mean absolute error of 5-fold cross validation is 0.994, 0.017, and 0.063 respectively. Interpretability study for the fine-tuned model is performed to enhance its transparency, trustworthiness, and physical interpretability. Moreover, the high-precision fine-tuned model is employed to generate 3000 samples. The proposed pre-trained model framework offers an efficient and robust means of accelerating design process and strengthening structural safety of underwater composite cylindrical shells, with immediate potential for extension to wider engineering applications.